The Great AI Cost Paradox

“America leads the frontier. China challenges the price of intelligence.”

Two developments reported on October 28 expose a fundamental contradiction in China's AI strategy: its model developers are competing aggressively on inference costs, while its domestic semiconductor industry faces substantial manufacturing disadvantages. At the same time, Beijing is considering restrictions that could limit the international diffusion of its most advanced AI technologies.

A third development, announced by Oracle on October 27, shows American infrastructure providers are increasingly competing on inference economics themselves. Together, these refine The Velocity Edge thesis: America leads the frontier, China challenges the economics of intelligence, and both are increasingly competing to control where the resulting economic value accumulates.

Reuters reported that Chinese memory manufacturer CXMT is preparing limited production of HBM3E chips within approximately two months. CXMT's chips reportedly deliver performance closer to SK Hynix's previous-generation HBM3 than its HBM3E products, while costing more than comparable foreign alternatives. These figures come from Reuters' industry sources rather than audited disclosures.

This is a consequential finding. China can produce increasingly competitive AI models without possessing an equally competitive domestic semiconductor manufacturing system. Low-cost inference software does not automatically translate into inexpensive sovereign computing infrastructure. Indeed, China's drive for technological independence may initially increase the capital required to deploy AI at scale.

Reuters also analyzed China's proposed AI technology export controls. The draft measures, released October 16, would distinguish between selling access to Chinese AI services internationally and transferring the underlying models, software or technologies to foreign operators. They cover 23 AI-related technologies and could affect models meeting specified computing, parameter-count or performance thresholds. Certain technology-transfer registration requirements took effect October 16; public comments remain open until November 15.

The potential reversal is substantial. China's advantage has partly rested on distributing accessible models that challenge American pricing. Beijing is now considering an architecture that would preserve domestic control over strategically valuable assets while allowing foreign customers to purchase access — strengthening intellectual-property protection and recurring revenues, but potentially weakening the open-weight diffusion strategy behind China's international competitiveness.

On October 27, Oracle announced plans to expand its Nvidia infrastructure while introducing accelerators from Google, Cerebras and SambaNova, and to substitute some Nvidia networking components with its own. Its Nvidia agreement encompasses expansion toward multi-gigawatt-scale Vera Rubin infrastructure, while the alternative accelerators serve different workloads, particularly inference. Financial terms were not disclosed.

The implication challenges a simplistic reading of the U.S.–China divide. America is not merely defending expensive frontier computing; its own companies are increasingly competing to reduce the cost of deploying intelligence. This may pressure Nvidia's ecosystem economics while expanding the market for inference services.

The mechanism is clear: hardware restrictions → domestic substitution → higher manufacturing costs → software optimization → lower inference prices → greater deployment demand. The unresolved question is whether additional workload volumes and productivity gains can compensate for the higher capital costs of domestic infrastructure.

This creates a potentially important rotation. Investors should distinguish companies earning scarcity premiums from those financing expensive technological substitution, and both from enterprises capturing productivity benefits through deployment. The AI Supercycle may keep expanding even while profitability diverges sharply across its infrastructure, software and industrial layers.

The Velocity Edge — Five Investment Implications

what to watch

• Chinese semiconductors: monitor CXMT's yields, costs and adoption — substitution must eventually become economically competitive.
• American infrastructure: evaluate Nvidia's pricing power against alternative inference accelerators, including Google's TPUs and Cerebras systems.
• Chinese frontier models: assess whether export restrictions reduce open-weight adoption or create stronger recurring revenues through controlled access.
• Industrial AI: prioritize measurable cost reductions and productivity gains over model benchmark improvements alone.
• Capital allocation: distinguish strategic investments in technological independence from projects capable of generating commercially attractive returns.

America leads the frontier. China is challenging the price of intelligence. The next financial battle will be won by those who can reduce the total cost of producing it — and capture the profits from its global deployment.